fork/minimap2
1
0
Fork 0
mirror of https://github.com/lh3/minimap2.git synced 2026-09-15 13:07:55 +08:00
Code Issues Packages Projects Releases Wiki Activity

avx2 support for chaining and alignment

Browse Source
This commit is contained in:
Saurabh
2021-06-26 10:08:37 -07:00
committed by Heng Li
parent a9ac74ffe1
commit 448341c96c
4 changed files with 679 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
align.c
View File

@@ -367,8 +367,8 @@ static void mm_align_pair(void *km, const mm_mapopt_t *opt, int qlen, const uint
#ifdef ALIGN_AVX
#ifdef __AVX512BW__
ksw_extd2_avx512(km, qlen, qseq, tlen, tseq, 5, mat, opt->q, opt->e, opt->q2, opt->e2, w, zdrop, end_bonus, flag, ez);
#else
ksw_extd2_sse(km, qlen, qseq, tlen, tseq, 5, mat, opt->q, opt->e, opt->q2, opt->e2, w, zdrop, end_bonus, flag, ez);
#elif __AVX2__
ksw_extd2_avx2(km, qlen, qseq, tlen, tseq, 5, mat, opt->q, opt->e, opt->q2, opt->e2, w, zdrop, end_bonus, flag, ez);
#endif
#else
ksw_extd2_sse(km, qlen, qseq, tlen, tseq, 5, mat, opt->q, opt->e, opt->q2, opt->e2, w, zdrop, end_bonus, flag, ez);

2
ext/TAL

Submodule ext/TAL updated: 75ce461846...4c317716e0

674
ksw2_extd2_avx.c
View File

@@ -29,7 +29,7 @@ Modified Copyright (C) 2021 Intel Corporation
*/
#include "ksw2_extd2_avx.h"
#ifdef __AVX512BW__
#ifdef __AVX512BW__
void ksw_extd2_avx512(void *km, int qlen, const uint8_t *query, int tlen, const uint8_t *target, int8_t m, const int8_t *mat,
int8_t q, int8_t e, int8_t q2, int8_t e2, int w, int zdrop, int end_bonus, int flag, ksw_extz_t *ez)
{
@@ -662,3 +662,675 @@ void ksw_extd2_avx512(void *km, int qlen, const uint8_t *query, int tlen, const
}
#endif
#ifdef __AVX2__
void ksw_extd2_avx2(void *km, int qlen, const uint8_t *query, int tlen, const uint8_t *target, int8_t m, const int8_t *mat,
int8_t q, int8_t e, int8_t q2, int8_t e2, int w, int zdrop, int end_bonus, int flag, ksw_extz_t *ez)
{
__m256i bt32_ = _mm256_setr_epi32(0,0,0,0,4,4,4,4);//8,8,8,8,12,12,12,12);
int8_t index[32] __attribute((aligned(64)));
for (int i=0; i<32; i++)
index[i] = i%16 - 1;
index[0] = 15;
index[16] = 31;
// index[32] = 47;
// index[48] = 63;
__m256i shf256a, shf256b, slli256;
__m256i ind256_slli = _mm256_load_si256((__m256i*) index);
__mmask8 mska = 0x00;//0x90
__mmask32 mskb = 0x10000;
__mmask32 mskc = 0x1;
__m256i mskb_v = _mm256_set_epi32(0,0,0,255,0,0,0,0);
__m256i mskc_v = _mm256_set_epi32(0,0,0,0,0,0,0,255);
__mmask32 mskc_ar[2] = {0x1, 0x10000};
__m256i mskc_ar_v[2];// = {0x1, 0x10000};
mskc_ar_v[0] = _mm256_set_epi32(0,0,0,0,0,0,0,255);
mskc_ar_v[1] = _mm256_set_epi32(0,0,0,255,0,0,0,0);
#define __dp_code_block1_pcl \
/* __mmask32 mskc_ = (t == st_) ? mskc_ar[(st0 - t*32)/16]:mskc; */ \
__m256i mskc_ = (t == st_) ? mskc_ar_v[(st0 - t*32)/16]:mskc_v; \
z = _mm256_load_si256(&s[t]); \
xt1 = _mm256_load_si256(&x[t]); /* xt1 <- x[r-1][t..t+15] */ \
/* tmp = _mm_srli_si128(xt1, 15); */ /* tmp <- x[r-1][t+15] */ \
tmp = _mm256_set1_epi8(((int8_t*)x)[t*32 + 31]); \
/* xt1 = _mm_or_si128(_mm_slli_si128(xt1, 1), x1_);*/ /* xt1 <- x[r-1][t-1..t+14] */ \
shf256a = _mm256_shuffle_epi8(xt1, ind256_slli); \
/*shf256b = _mm256_shuffle_i32x4(shf256a, shf256a, 0x00); */ \
shf256b = _mm256_permute2x128_si256(shf256a, shf256a, 0); \
/* slli256 = _mm256_mask_blend_epi8(mskb, shf256a, shf256b); */ \
slli256 = _mm256_blendv_epi8(shf256a, shf256b, mskb_v); \
xt1 = _mm256_blendv_epi8(slli256, x1_, mskc_); \
x1_ = tmp; \
vt1 = _mm256_load_si256(&v[t]); /* vt1 <- v[r-1][t..t+15] */ \
/* tmp = _mm_srli_si128(vt1, 15); */ /* tmp <- v[r-1][t+15] */ \
tmp = _mm256_set1_epi8(((int8_t*)v)[t*32 + 31]); \
/* vt1 = _mm_or_si128(_mm_slli_si128(vt1, 1), v1_); *//* vt1 <- v[r-1][t-1..t+14] */ \
shf256a = _mm256_shuffle_epi8(vt1, ind256_slli); \
/*shf256b = _mm256_shuffle_i32x4(shf256a, shf256a, 0x00); */ \
shf256b = _mm256_permute2x128_si256(shf256a, shf256a, 0); \
/* slli256 = _mm256_mask_blend_epi8(mskb, shf256a, shf256b); */ \
slli256 = _mm256_blendv_epi8(shf256a, shf256b, mskb_v); \
vt1 = _mm256_blendv_epi8(slli256, v1_, mskc_); \
v1_ = tmp; \
a = _mm256_add_epi8(xt1, vt1); /* a <- x[r-1][t-1..t+14] + v[r-1][t-1..t+14] */ \
ut = _mm256_load_si256(&u[t]); /* ut <- u[t..t+15] */ \
b = _mm256_add_epi8(_mm256_load_si256(&y[t]), ut); /* b <- y[r-1][t..t+15] + u[r-1][t..t+15] */ \
x2t1= _mm256_load_si256(&x2[t]); \
/* tmp = _mm_srli_si128(x2t1, 15);*/ \
tmp = _mm256_set1_epi8(((int8_t*)x2)[t*32 + 31]); \
/* x2t1= _mm_or_si128(_mm_slli_si128(x2t1, 1), x21_); */ \
shf256a = _mm256_shuffle_epi8(x2t1, ind256_slli); \
/*shf256b = _mm256_shuffle_i32x4(shf256a, shf256a, 0x00);*/ \
shf256b = _mm256_permute2x128_si256(shf256a, shf256a, 0); \
/* slli256 = _mm256_mask_blend_epi8(mskb, shf256a, shf256b); */ \
slli256 = _mm256_blendv_epi8(shf256a, shf256b, mskb_v); \
x2t1 = _mm256_blendv_epi8(slli256, x21_, mskc_); \
x21_= tmp; \
a2= _mm256_add_epi8(x2t1, vt1); \
b2= _mm256_add_epi8(_mm256_load_si256(&y2[t]), ut);
#define __dp_code_block2_pcl \
_mm256_storeu_si256(&u[t], _mm256_sub_epi8(z, vt1)); /* u[r][t..t+15] <- z - v[r-1][t-1..t+14] */ \
_mm256_storeu_si256(&v[t], _mm256_sub_epi8(z, ut)); /* v[r][t..t+15] <- z - u[r-1][t..t+15] */ \
tmp = _mm256_sub_epi8(z, q_); \
a = _mm256_sub_epi8(a, tmp); \
b = _mm256_sub_epi8(b, tmp); \
tmp = _mm256_sub_epi8(z, q2_); \
a2= _mm256_sub_epi8(a2, tmp); \
b2= _mm256_sub_epi8(b2, tmp);
// __mmask64 msk_ar[5] = {0xFFFF, 0xFFFFFFFF, 0xFFFFFFFFFFFF, 0xFFFFFFFFFFFFFFFF, 0xFFFFFFFFFFFFFFFF};
__mmask32 msk_ar2[3] = {0xFFFF, 0xFFFF, 0xFFFFFFFF};
__m256i msk_ar2_v[3];
msk_ar2_v[0] = _mm256_set_epi32(0,0,0,0,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF);
msk_ar2_v[1] = _mm256_set_epi32(0,0,0,0,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF);
msk_ar2_v[2] = _mm256_set_epi32(0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF);
int r, t, qe = q + e, n_col_, *off = 0, *off_end = 0, tlen_, qlen_, last_st, last_en, wl, wr, max_sc, min_sc, long_thres, long_diff;
int with_cigar = !(flag&KSW_EZ_SCORE_ONLY), approx_max = !!(flag&KSW_EZ_APPROX_MAX);
int32_t *H = 0, H0 = 0, last_H0_t = 0;
uint8_t *qr, *sf, *mem, *mem2 = 0;
__m256i q_, q2_, qe_, qe2_, zero_, sc_mch_, sc_mis_, m1_, sc_N_;
__m256i *u, *v, *x, *y, *x2, *y2, *s, *p = 0;
__m256i one_, two_, three_, four_, s1_, s2_, s3_, s4_;
ksw_reset_extz(ez);
if (m <= 1 || qlen <= 0 || tlen <= 0) return;
if (q2 + e2 < q + e) t = q, q = q2, q2 = t, t = e, e = e2, e2 = t; // make sure q+e no larger than q2+e2
s1_ = _mm256_set1_epi8(0x08);
s2_ = _mm256_set1_epi8(0x10);
s3_ = _mm256_set1_epi8(0x20);
s4_ = _mm256_set1_epi8(0x40);
one_ = _mm256_set1_epi8(1);
two_ = _mm256_set1_epi8(2);
three_ = _mm256_set1_epi8(3);
four_ = _mm256_set1_epi8(4);
zero_ = _mm256_set1_epi8(0);
q_ = _mm256_set1_epi8(q);
q2_ = _mm256_set1_epi8(q2);
qe_ = _mm256_set1_epi8(q + e);
qe2_ = _mm256_set1_epi8(q2 + e2);
sc_mch_ = _mm256_set1_epi8(mat[0]);
sc_mis_ = _mm256_set1_epi8(mat[1]);
sc_N_ = mat[m*m-1] == 0? _mm256_set1_epi8(-e2) : _mm256_set1_epi8(mat[m*m-1]);
m1_ = _mm256_set1_epi8(m - 1); // wildcard
if (w < 0) w = tlen > qlen? tlen : qlen;
wl = wr = w;
tlen_ = (tlen + 31) / 32;
n_col_ = qlen < tlen? qlen : tlen;
n_col_ = ((n_col_ < w + 1? n_col_ : w + 1) + 31) / 32 + 1;
qlen_ = (qlen + 31) / 32;
for (t = 1, max_sc = mat[0], min_sc = mat[1]; t < m * m; ++t) {
max_sc = max_sc > mat[t]? max_sc : mat[t];
min_sc = min_sc < mat[t]? min_sc : mat[t];
}
if (-min_sc > 2 * (q + e)) return; // otherwise, we won't see any mismatches
long_thres = e != e2? (q2 - q) / (e - e2) - 1 : 0;
if (q2 + e2 + long_thres * e2 > q + e + long_thres * e)
++long_thres;
long_diff = long_thres * (e - e2) - (q2 - q) - e2;
mem = (uint8_t*)kcalloc(km, tlen_ * 8 + qlen_ + 1 + 63, 64);
u = (__m256i*)(((size_t)mem + 31) >> 5 << 5); // 16-byte aligned
v = u + tlen_, x = v + tlen_, y = x + tlen_, x2 = y + tlen_, y2 = x2 + tlen_;
s = y2 + tlen_, sf = (uint8_t*)(s + tlen_), qr = sf + tlen_ * 32;
memset(u, -q - e, tlen_ * 32);
memset(v, -q - e, tlen_ * 32);
memset(x, -q - e, tlen_ * 32);
memset(y, -q - e, tlen_ * 32);
memset(x2, -q2 - e2, tlen_ * 32);
memset(y2, -q2 - e2, tlen_ * 32);
if (!approx_max) {
H = (int32_t*)kmalloc(km, tlen_ * 32 * 4);
for (t = 0; t < tlen_ * 32; ++t) H[t] = KSW_NEG_INF;
}
if (with_cigar) {
mem2 = (uint8_t*)kmalloc(km, ((size_t)(qlen + tlen - 1) * n_col_ + 1) * 32);
p = (__m256i*)(((size_t)mem2 + 31) >> 5 << 5);
off = (int*)kmalloc(km, (qlen + tlen - 1) * sizeof(int) * 2);
off_end = off + qlen + tlen - 1;
}
for (t = 0; t < qlen; ++t) qr[t] = query[qlen - 1 - t];
memcpy(sf, target, tlen);
for (r = 0, last_st = last_en = -1; r < qlen + tlen - 1; ++r) {
int st = 0, en = tlen - 1, st0, en0, st_, en_;
int8_t x1, x21, v1;
uint8_t *qrr = qr + (qlen - 1 - r);
int8_t *u8 = (int8_t*)u, *v8 = (int8_t*)v, *x8 = (int8_t*)x, *x28 = (int8_t*)x2;
__m256i x1_, x21_, v1_;
// find the boundaries
if (st < r - qlen + 1) st = r - qlen + 1;
if (en > r) en = r;
if (st < (r-wr+1)>>1) st = (r-wr+1)>>1; // take the ceil
if (en > (r+wl)>>1) en = (r+wl)>>1; // take the floor
if (st > en) {
ez->zdropped = 1;
break;
}
st0 = st, en0 = en;
int st_new = st / 16 * 16, en_new = (en + 16) / 16 * 16 - 1;
// int st_new = st / 64 * 64, en_new = (en + 64) / 64 * 64 - 1;
int stb = st, enb = en;
st = st / 32 * 32, en = (en + 32) / 32 * 32 - 1;
int stn = stb / 16 * 16, enn = (enb + 16) / 16 * 16 - 1;
// set boundary conditions
if (st_new > 0) {
if (st_new - 1 >= last_st && st_new - 1 <= last_en) {
x1 = x8[st_new - 1], x21 = x28[st_new - 1], v1 = v8[st_new - 1]; // (r-1,s-1) calculated in the last round
} else {
x1 = -q - e, x21 = -q2 - e2;
v1 = -q - e;
}
} else {
x1 = -q - e, x21 = -q2 - e2;
v1 = r == 0? -q - e : r < long_thres? -e : r == long_thres? long_diff : -e2;
}
if (en_new >= r) {
((int8_t*)y)[r] = -q - e, ((int8_t*)y2)[r] = -q2 - e2;
u8[r] = r == 0? -q - e : r < long_thres? -e : r == long_thres? long_diff : -e2;
}
// loop fission: set scores first
if (!(flag & KSW_EZ_GENERIC_SC)) {
for (t = st0; t <= en0; t += 32) {
__m256i sq, st, tmp_256, mask_256;
__mmask32 tmp, mask;
sq = _mm256_loadu_si256((__m256i*)&sf[t]);
st = _mm256_loadu_si256((__m256i*)&qrr[t]);
// mask = (_mm256_cmpeq_epi8_mask(sq, m1_) | _mm256_cmpeq_epi8_mask(st, m1_));
mask_256 = _mm256_or_si256(_mm256_cmpeq_epi8(sq, m1_), _mm256_cmpeq_epi8(st, m1_));
tmp_256 = _mm256_cmpeq_epi8(sq, st);
tmp_256 = _mm256_blendv_epi8(sc_mis_, sc_mch_, tmp_256);
tmp_256 = _mm256_blendv_epi8(tmp_256, sc_N_, mask_256);
if (t + 32 > en0)
{
__mmask32 msk;
int ind = (en0 - t + 16)/16;
assert(ind >= 0 && ind < 3);
msk = msk_ar2[ind];
__m256i msk_v = msk_ar2_v[ind];
__m256i str =_mm256_and_si256( msk_v, tmp_256);// msk_ar2_v[ind];
_mm256_storeu_si256((__m256i*)((int8_t*)s + t), str);
}
else
_mm256_storeu_si256((__m256i*)((int8_t*)s + t), tmp_256);
}
} else {
for (t = st0; t <= en0; ++t)
((uint8_t*)s)[t] = mat[sf[t] * m + qrr[t]];
}
// core loop
// fprintf(stderr, "- r: %d, x1: %d, x21: %d, v1: %d, en_new: %d, e: %d, q: %d\n",
//r, x1, x21, v1, en_new, e, q);
x1_ = _mm256_set1_epi8((uint8_t)x1);
x21_ = _mm256_set1_epi8((uint8_t)x21);
v1_ = _mm256_set1_epi8((uint8_t)v1);
//st_ = st / 16, en_ = en / 16;
st_ = st / 32, en_ = en / 32;
//assert(en_ - st_ + 1 <= n_col_);
if (!with_cigar) { // score only
for (t = st_; t <= en_; ++t) {
__m256i z, a, b, a2, b2, xt1, x2t1, vt1, ut, tmp;
__dp_code_block1_pcl;
z = _mm256_max_epi8(z, a);
z = _mm256_max_epi8(z, b);
z = _mm256_max_epi8(z, a2);
z = _mm256_max_epi8(z, b2);
z = _mm256_min_epi8(z, sc_mch_);
// __dp_code_block2_pcl; // save u[] and v[]; update a, b, a2 and b2
if (t == en_) {
__mmask32 msk;
int ind = (en0 - t*32 + 16)/16;//doubt
msk = msk_ar2[ind];
// fprintf(stderr, "en0: %d, t: %d, ind: %d, msk: %d\n", en0, t, ind, msk);
_mm256_storeu_si256(&u[t], _mm256_and_si256(msk_ar2_v[ind], _mm256_sub_epi8(z, vt1)));
_mm256_storeu_si256(&v[t], _mm256_and_si256(msk_ar2_v[ind], _mm256_sub_epi8(z, ut)));
tmp = _mm256_sub_epi8(z, q_);
a = _mm256_sub_epi8(a, tmp);
b = _mm256_sub_epi8(b, tmp);
tmp = _mm256_sub_epi8(z, q2_);
a2= _mm256_sub_epi8(a2, tmp);
b2= _mm256_sub_epi8(b2, tmp);
}
else {
_mm256_storeu_si256(&u[t], _mm256_sub_epi8(z, vt1));
_mm256_storeu_si256(&v[t], _mm256_sub_epi8(z, ut));
tmp = _mm256_sub_epi8(z, q_);
a = _mm256_sub_epi8(a, tmp);
b = _mm256_sub_epi8(b, tmp);
tmp = _mm256_sub_epi8(z, q2_);
a2= _mm256_sub_epi8(a2, tmp);
b2= _mm256_sub_epi8(b2, tmp);
}
if (t == en_) {
__mmask32 msk;
int ind = (en0 - t*32 + 16)/16;//doubt
//assert(ind >= 0);
msk = msk_ar2[ind];
// fprintf(stderr, "en0: %d, t: %d, ind: %d, msk: %d\n", en0, t, ind, msk);
_mm256_storeu_si256(&x[t], _mm256_and_si256(msk_ar2_v[ind], _mm256_sub_epi8(_mm256_max_epi8(a, zero_), qe_)));
_mm256_storeu_si256(&y[t], _mm256_and_si256(msk_ar2_v[ind], _mm256_sub_epi8(_mm256_max_epi8(b, zero_), qe_)));
_mm256_storeu_si256(&x2[t], _mm256_and_si256(msk_ar2_v[ind], _mm256_sub_epi8(_mm256_max_epi8(a2, zero_), qe2_)));
_mm256_storeu_si256(&y2[t], _mm256_and_si256(msk_ar2_v[ind], _mm256_sub_epi8(_mm256_max_epi8(b2, zero_), qe2_)));
}
else
{
_mm256_storeu_si256(&x[t], _mm256_sub_epi8(_mm256_max_epi8(a, zero_), qe_));
_mm256_storeu_si256(&y[t], _mm256_sub_epi8(_mm256_max_epi8(b, zero_), qe_));
_mm256_storeu_si256(&x2[t], _mm256_sub_epi8(_mm256_max_epi8(a2, zero_), qe2_));
_mm256_storeu_si256(&y2[t], _mm256_sub_epi8(_mm256_max_epi8(b2, zero_), qe2_));
}
// for (int l=0; l<64; l++)
// fprintf(stderr, "%d ", ((int8_t*)x)[l]);
}
} else if (!(flag&KSW_EZ_RIGHT)) { // gap left-alignment
__m256i *pr = p + (size_t)r * n_col_ - st_;
off[r] = st, off_end[r] = en;
for (t = st_; t < en_; ++t) {
__m256i d, z, a, b, a2, b2, xt1, x2t1, vt1, ut, tmp;
// __mmask32 tmp_mask;
__m256i tmp_mask;
__dp_code_block1_pcl;
d = _mm256_blendv_epi8(zero_, one_, _mm256_cmpgt_epi8(a, z)); // d = a > z? 1 : 0
z = _mm256_max_epi8(z, a);
d = _mm256_blendv_epi8(d, two_,_mm256_cmpgt_epi8(b, z)); // d = b > z? 2 : d
z = _mm256_max_epi8(z, b);
d = _mm256_blendv_epi8(d, three_,_mm256_cmpgt_epi8(a2, z)); // d = a2 > z? 3 : d
z = _mm256_max_epi8(z, a2);
d = _mm256_blendv_epi8(d, four_,_mm256_cmpgt_epi8(b2, z)); // d = b2 > z? 4 : d
z = _mm256_max_epi8(z, b2);
z = _mm256_min_epi8(z, sc_mch_);
// __dp_code_block2_pcl;
_mm256_storeu_si256(&u[t], _mm256_sub_epi8(z, vt1));
_mm256_storeu_si256(&v[t], _mm256_sub_epi8(z, ut));
tmp = _mm256_sub_epi8(z, q_);
a = _mm256_sub_epi8(a, tmp);
b = _mm256_sub_epi8(b, tmp);
tmp = _mm256_sub_epi8(z, q2_);
a2= _mm256_sub_epi8(a2, tmp);
b2= _mm256_sub_epi8(b2, tmp);
tmp_mask = _mm256_cmpgt_epi8(a, zero_);
_mm256_storeu_si256(&x[t], _mm256_sub_epi8(_mm256_blendv_epi8(zero_, a, tmp_mask), qe_));
d = _mm256_or_si256(d, _mm256_blendv_epi8( zero_, s1_, tmp_mask)); // d = a > 0? 1<<3 : 0
tmp_mask = _mm256_cmpgt_epi8(b, zero_);
_mm256_storeu_si256(&y[t], _mm256_sub_epi8(_mm256_blendv_epi8( zero_, b, tmp_mask), qe_));
d = _mm256_or_si256(d, _mm256_blendv_epi8(zero_, s2_, tmp_mask)); // d = b > 0? 1<<4 : 0
tmp_mask = _mm256_cmpgt_epi8(a2, zero_);
_mm256_storeu_si256(&x2[t], _mm256_sub_epi8(_mm256_blendv_epi8(zero_, a2,tmp_mask), qe2_));
d = _mm256_or_si256(d, _mm256_blendv_epi8( zero_, s3_,tmp_mask)); // d = a > 0? 1<<5 : 0
tmp_mask = _mm256_cmpgt_epi8(b2, zero_);
_mm256_storeu_si256(&y2[t], _mm256_sub_epi8(_mm256_blendv_epi8(zero_, b2,tmp_mask), qe2_));
d = _mm256_or_si256(d, _mm256_blendv_epi8(zero_, s4_,tmp_mask)); // d = b > 0? 1<<6 : 0
_mm256_storeu_si256(&pr[t], d);
}
{
__m256i d, z, a, b, a2, b2, xt1, x2t1, vt1, ut, tmp;
//__mmask32 tmp_mask;
__m256i tmp_mask;
__dp_code_block1_pcl;
d = _mm256_blendv_epi8(zero_,one_, _mm256_cmpgt_epi8(a, z)); // d = a > z? 1 : 0
z = _mm256_max_epi8(z, a);
d = _mm256_blendv_epi8(d, two_, _mm256_cmpgt_epi8(b, z)); // d = b > z? 2 : d
z = _mm256_max_epi8(z, b);
d = _mm256_blendv_epi8(d, three_, _mm256_cmpgt_epi8(a2, z)); // d = a2 > z? 3 : d
z = _mm256_max_epi8(z, a2);
d = _mm256_blendv_epi8(d, four_, _mm256_cmpgt_epi8(b2, z)); // d = b2 > z? 3 : d
z = _mm256_max_epi8(z, b2);
z = _mm256_min_epi8(z, sc_mch_);
// __dp_code_block2_pcl;
{
__mmask32 msk;
int ind = (en0 - t*32 + 16)/16;//doubt
// //assert(ind >= 0 && ind < 5);
msk = msk_ar2[ind];
_mm256_storeu_si256(&u[t], _mm256_and_si256(msk_ar2_v[ind], _mm256_sub_epi8(z, vt1)));
_mm256_storeu_si256(&v[t], _mm256_and_si256(msk_ar2_v[ind], _mm256_sub_epi8(z, ut)));
tmp = _mm256_sub_epi8(z, q_);
a = _mm256_sub_epi8(a, tmp);
b = _mm256_sub_epi8(b, tmp);
tmp = _mm256_sub_epi8(z, q2_);
a2= _mm256_sub_epi8(a2, tmp);
b2= _mm256_sub_epi8(b2, tmp);
}
{
__mmask32 msk;
int ind = (en0 - t*32 + 16)/16;//doubt
msk = msk_ar2[ind];
__m256i msk_v= msk_ar2_v[ind];
off_end[r] -= (2-ind)*16;//doubt
tmp_mask = _mm256_cmpgt_epi8(a, zero_);
_mm256_storeu_si256(&x[t], _mm256_and_si256(msk_v, _mm256_sub_epi8(_mm256_blendv_epi8(zero_, a, tmp_mask), qe_)));
d = _mm256_or_si256(d, _mm256_blendv_epi8(zero_, s1_, tmp_mask)); // d = a > 0? 1<<3 : 0
tmp_mask = _mm256_cmpgt_epi8(b, zero_);
_mm256_storeu_si256(&y[t], _mm256_and_si256(msk_v , _mm256_sub_epi8(_mm256_blendv_epi8(zero_, b, tmp_mask), qe_)));
d = _mm256_or_si256(d, _mm256_blendv_epi8(zero_, s2_, tmp_mask)); // d = b > 0? 1<<4 : 0
tmp_mask = _mm256_cmpgt_epi8(a2, zero_);
_mm256_storeu_si256(&x2[t], _mm256_and_si256(msk_v, _mm256_sub_epi8(_mm256_blendv_epi8(zero_, a2, tmp_mask), qe2_)));
d = _mm256_or_si256(d, _mm256_blendv_epi8(zero_, s3_, tmp_mask)); // d = a > 0? 1<<5 : 0
tmp_mask = _mm256_cmpgt_epi8(b2, zero_);
_mm256_storeu_si256(&y2[t], _mm256_and_si256(msk_v, _mm256_sub_epi8(_mm256_blendv_epi8(zero_, b2, tmp_mask), qe2_)));
d = _mm256_or_si256(d, _mm256_blendv_epi8(zero_, s4_, tmp_mask)); // d = b > 0? 1<<6 : 0
_mm256_storeu_si256(&pr[t], _mm256_and_si256(msk_v, d));
//_mm256_mask_storeu_epi8(&pr[t], msk, d);
}
}
} else { // gap right-alignment
__m256i *pr = p + (size_t)r * n_col_ - st_;
off[r] = st, off_end[r] = en;
// off[r] = stn, off_end[r] = enn;
// fprintf(stderr, "t: %d, st0: %d\n", st_, st0);
for (t = st_; t < en_; ++t) {
__m256i d, z, a, b, a2, b2, xt1, x2t1, vt1, ut, tmp;
//__mmask32 tmp_mask;
__m256i tmp_mask;
__dp_code_block1_pcl;
d = _mm256_blendv_epi8(one_, zero_, _mm256_cmpgt_epi8(z, a));
z = _mm256_max_epi8(z, a);
// d = _mm256_blendv_epi8(_mm256_cmpgt_epi8(z, b), _mm256_set1_epi8(2), d);
// d = z > b? d : 2
d = _mm256_blendv_epi8(two_, d,_mm256_cmpgt_epi8(z, b) ); // d = z > b? d : 2
z = _mm256_max_epi8(z, b);
// d = z > a2? d : 3
d = _mm256_blendv_epi8(three_, d, _mm256_cmpgt_epi8(z, a2)); // d = z > a2? d : 3
z = _mm256_max_epi8(z, a2);
// d = z > b2? d : 4
d = _mm256_blendv_epi8(four_, d, _mm256_cmpgt_epi8(z, b2)); // d = z > b2? d : 4
z = _mm256_max_epi8(z, b2);
z = _mm256_min_epi8(z, sc_mch_);
// __dp_code_block2_pcl;
__mmask32 msk;
{
_mm256_storeu_si256(&u[t], _mm256_sub_epi8(z, vt1));
_mm256_storeu_si256(&v[t], _mm256_sub_epi8(z, ut));
tmp = _mm256_sub_epi8(z, q_);
a = _mm256_sub_epi8(a, tmp);
b = _mm256_sub_epi8(b, tmp);
tmp = _mm256_sub_epi8(z, q2_);
a2= _mm256_sub_epi8(a2, tmp);
b2= _mm256_sub_epi8(b2, tmp);
}
{
tmp_mask = _mm256_cmpgt_epi8(zero_, a);
_mm256_storeu_si256(&x[t], _mm256_sub_epi8(_mm256_blendv_epi8(a, zero_,tmp_mask), qe_));
// d = a > 0? 1<<3 : 0
d = _mm256_or_si256(d, _mm256_blendv_epi8(s1_, zero_,tmp_mask)); // d = a > 0? 1<<3 : 0
tmp_mask = _mm256_cmpgt_epi8(zero_, b);
_mm256_storeu_si256(&y[t], _mm256_sub_epi8(_mm256_blendv_epi8(b, zero_,tmp_mask), qe_));
// d = b > 0? 1<<4 : 0
d = _mm256_or_si256(d, _mm256_blendv_epi8(s2_, zero_,tmp_mask)); // d = b > 0? 1<<4 : 0
tmp_mask = _mm256_cmpgt_epi8(zero_, a2);
_mm256_storeu_si256(&x2[t], _mm256_sub_epi8(_mm256_blendv_epi8(a2, zero_,tmp_mask), qe2_));
// d = a > 0? 1<<5 : 0
d = _mm256_or_si256(d, _mm256_blendv_epi8(s3_, zero_,tmp_mask)); // d = a > 0? 1<<5 : 0
tmp_mask = _mm256_cmpgt_epi8(zero_, b2);
_mm256_storeu_si256(&y2[t], _mm256_sub_epi8(_mm256_blendv_epi8(b2, zero_,tmp_mask), qe2_));
// d = b > 0? 1<<6 : 0
d = _mm256_or_si256(d, _mm256_blendv_epi8(s4_, zero_,tmp_mask)); // d = b > 0? 1<<6 : 0
_mm256_storeu_si256(&pr[t], d);
}
}
//for (t = st_; t <= en_; ++t)// Last iteration unrolled
{
__m256i d, z, a, b, a2, b2, xt1, x2t1, vt1, ut, tmp;
//__mmask32 tmp_mask;
__m256i tmp_mask;
__dp_code_block1_pcl;
d = _mm256_blendv_epi8(one_, zero_,_mm256_cmpgt_epi8(z, a) );
z = _mm256_max_epi8(z, a);
// d = z > b? d : 2
d = _mm256_blendv_epi8( two_, d, _mm256_cmpgt_epi8(z, b)); // d = z > b? d : 2
z = _mm256_max_epi8(z, b);
// d = z > a2? d : 3
d = _mm256_blendv_epi8( three_, d, _mm256_cmpgt_epi8(z, a2)); // d = z > a2? d : 3
z = _mm256_max_epi8(z, a2);
// d = z > b2? d : 4
d = _mm256_blendv_epi8(four_, d, _mm256_cmpgt_epi8(z, b2)); // d = z > b2? d : 4
z = _mm256_max_epi8(z, b2);
z = _mm256_min_epi8(z, sc_mch_);
// __dp_code_block2_pcl;
__mmask32 msk;
// __mmask64 msk;
int ind = (en0 - t*32 + 16)/16;//doubt
msk = msk_ar2[ind];
off_end[r] -= (2-ind)*16;//doubt
_mm256_storeu_si256(&u[t], _mm256_and_si256(msk_ar2_v[ind], _mm256_sub_epi8(z, vt1)));
_mm256_storeu_si256(&v[t], _mm256_and_si256(msk_ar2_v[ind], _mm256_sub_epi8(z, ut)));
tmp = _mm256_sub_epi8(z, q_);
a = _mm256_sub_epi8(a, tmp);
b = _mm256_sub_epi8(b, tmp);
tmp = _mm256_sub_epi8(z, q2_);
a2= _mm256_sub_epi8(a2, tmp);
b2= _mm256_sub_epi8(b2, tmp);
tmp_mask = _mm256_cmpgt_epi8(zero_, a);
_mm256_storeu_si256(&x[t], _mm256_and_si256(msk_ar2_v[ind], _mm256_sub_epi8(_mm256_blendv_epi8(a, zero_, tmp_mask), qe_)));
d = _mm256_or_si256(d, _mm256_blendv_epi8( s1_, zero_, tmp_mask)); // d = a > 0? 1<<3 : 0
tmp_mask = _mm256_cmpgt_epi8(zero_, b);
_mm256_storeu_si256(&y[t], _mm256_and_si256(msk_ar2_v[ind], _mm256_sub_epi8(_mm256_blendv_epi8(b, zero_,tmp_mask), qe_)));
d = _mm256_or_si256(d, _mm256_blendv_epi8(s2_, zero_,tmp_mask)); // d = b > 0? 1<<4 : 0
tmp_mask = _mm256_cmpgt_epi8(zero_, a2);
_mm256_storeu_si256(&x2[t], _mm256_and_si256(msk_ar2_v[ind], _mm256_sub_epi8(_mm256_blendv_epi8(a2, zero_,tmp_mask), qe2_)));
d = _mm256_or_si256(d, _mm256_blendv_epi8(s3_, zero_,tmp_mask)); // d = a > 0? 1<<5 : 0
tmp_mask = _mm256_cmpgt_epi8(zero_, b2);
_mm256_storeu_si256(&y2[t], _mm256_and_si256(msk_ar2_v[ind], _mm256_sub_epi8(_mm256_blendv_epi8(b2, zero_,tmp_mask), qe2_)));
d = _mm256_or_si256(d, _mm256_blendv_epi8(s4_, zero_, tmp_mask)); // d = b > 0? 1<<6 : 0
// _mm256_storeu_si256(&pr[t], d);
_mm256_storeu_si256(&pr[t], _mm256_and_si256(msk_ar2_v[ind], d));
}
}
if (!approx_max) { // find the exact max with a 32-bit score array
int32_t max_H, max_t;
// compute H[], max_H and max_t
if (r > 0) {
int32_t HH[8], tt[8], en1 = st0 + (en0 - st0) / 8 * 8, i; //doubt
__m256i max_H_, max_t_;
max_H = H[en0] = en0 > 0? H[en0-1] + u8[en0] : H[en0] + v8[en0]; // special casing the last element
max_t = en0;
max_H_ = _mm256_set1_epi32(max_H);
max_t_ = _mm256_set1_epi32(max_t);
for (t = st0; t < en1; t += /*4*/8) { // this implements: H[t]+=v8[t]-qe; if(H[t]>max_H) max_H=H[t],max_t=t;
__m256i H1, tmp, t_;
H1 = _mm256_loadu_si256((__m256i*)&H[t]);
//__m128i t__ = _mm_load_si128((__m128i*) &v8[t]);
//t_ = _mm256_cvtepi8_epi32(t__);
t_ = _mm256_setr_epi32(v8[t], v8[t+1], v8[t+2], v8[t+3], v8[t+4], v8[t+5], v8[t+6], v8[t+7]);
H1 = _mm256_add_epi32(H1, t_);
_mm256_storeu_si256((__m256i*)&H[t], H1);
// making it 4 lanes to match accuracy
__m256i shfH, shft, max1, max2;
__mmask8 tmp_mask;
__m256i tmp_mask_v;
t_ = _mm256_set1_epi32(t);
t_ = _mm256_add_epi32(t_, bt32_);
//shfH = _mm256_shuffle_i32x4(H1, H1, 0x1);//doubt
//shft = _mm256_shuffle_i32x4(t_, t_, 0x1);//doubt
shfH = _mm256_permute2x128_si256(H1, H1, 1);
shft = _mm256_permute2x128_si256(t_, t_, 1);
tmp_mask_v = _mm256_cmpgt_epi32(shfH, H1);
max1 = _mm256_or_si256(_mm256_and_si256(tmp_mask_v, shfH) , _mm256_andnot_si256(tmp_mask_v, H1));
max2 = _mm256_or_si256(_mm256_and_si256(tmp_mask_v, shft) , _mm256_andnot_si256(tmp_mask_v, t_));
/* //--shfH = _mm256_shuffle_i32x4(max1, max1, 0x2);//doubt
//--shft = _mm256_shuffle_i32x4(max2, max2, 0x2);//doubt
//--tmp_mask = _mm256_cmpgt_epi32_mask(shfH, max1);
//-max1 = _mm256_mask_blend_epi32(tmp_mask, max1, shfH);
//--max2 = _mm256_mask_blend_epi32(tmp_mask, max2, shft);
*/
// tmp_mask = _mm256_cmpgt_epi32_mask(max1, max_H_);
// max_H_ = _mm256_mask_blend_epi32(tmp_mask, max_H_, max1);
// max_t_ = _mm256_mask_blend_epi32(tmp_mask, max_t_, max2);
tmp_mask_v = _mm256_cmpgt_epi32(max1, max_H_);
max_H_ = _mm256_or_si256(_mm256_and_si256(tmp_mask_v, max1) , _mm256_andnot_si256(tmp_mask_v, max_H_));
max_t_ = _mm256_or_si256(_mm256_and_si256(tmp_mask_v, max2) , _mm256_andnot_si256(tmp_mask_v, max_t_));
}
_mm256_storeu_si256((__m256i*)HH, max_H_);
_mm256_storeu_si256((__m256i*)tt, max_t_);
int rem = (en0 - t) / 4;
for (int l=0; l<rem; l++) {
int bt = t;
for (int j=0; j<4; j++) {
H[t] += (int32_t)v8[t];
if (H[t] > HH[j]) {
HH[j] = H[t];
tt[j] = bt;
}
t++;
}
}
for (i = 0; i < 4; ++i)
if (max_H < HH[i]) max_H = HH[i], max_t = tt[i] + i;
for (; t < en0; ++t) { // for the rest of values that haven't been computed with SSE
H[t] += (int32_t)v8[t];
if (H[t] > max_H) {
max_H = H[t], max_t = t;
}
}
} else H[0] = v8[0] - qe, max_H = H[0], max_t = 0; // special casing r==0
// update ez
if (en0 == tlen - 1 && H[en0] > ez->mte) {
ez->mte = H[en0], ez->mte_q = r - en_new;
}
if (r - st0 == qlen - 1 && H[st0] > ez->mqe) {
ez->mqe = H[st0], ez->mqe_t = st0;
}
if (ksw_apply_zdrop(ez, 1, max_H, r, max_t, zdrop, e2)) {
break;
}
if (r == qlen + tlen - 2 && en0 == tlen - 1) {
ez->score = H[tlen - 1];
}
} else { // find approximate max; Z-drop might be inaccurate, too.
if (r > 0) {
if (last_H0_t >= st0 && last_H0_t <= en0 && last_H0_t + 1 >= st0 && last_H0_t + 1 <= en0) {
int32_t d0 = v8[last_H0_t];
int32_t d1 = u8[last_H0_t + 1];
if (d0 > d1) H0 += d0;
else H0 += d1, ++last_H0_t;
} else if (last_H0_t >= st0 && last_H0_t <= en0) {
H0 += v8[last_H0_t];
} else {
++last_H0_t, H0 += u8[last_H0_t];
}
} else H0 = v8[0] - qe, last_H0_t = 0;
if ((flag & KSW_EZ_APPROX_DROP) && ksw_apply_zdrop(ez, 1, H0, r, last_H0_t, zdrop, e2)) {
break;
}
if (r == qlen + tlen - 2 && en0 == tlen - 1) {
ez->score = H0;
}
}
// last_st = st, last_en = en;
last_st = st_new, last_en = en_new;
}
kfree(km, mem);
if (!approx_max) kfree(km, H);
if (with_cigar) { // backtrack
int rev_cigar = !!(flag & KSW_EZ_REV_CIGAR);
if (!ez->zdropped && !(flag&KSW_EZ_EXTZ_ONLY)) {
ksw_backtrack(km, 1, rev_cigar, 0, (uint8_t*)p, off, off_end, n_col_*32, tlen-1, qlen-1, &ez->m_cigar, &ez->n_cigar, &ez->cigar);
} else if (!ez->zdropped && (flag&KSW_EZ_EXTZ_ONLY) && ez->mqe + end_bonus > (int)ez->max) {
ez->reach_end = 1;
ksw_backtrack(km, 1, rev_cigar, 0, (uint8_t*)p, off, off_end, n_col_*32, ez->mqe_t, qlen-1, &ez->m_cigar, &ez->n_cigar, &ez->cigar);
} else if (ez->max_t >= 0 && ez->max_q >= 0) {
ksw_backtrack(km, 1, rev_cigar, 0, (uint8_t*)p, off, off_end, n_col_*32, ez->max_t, ez->max_q, &ez->m_cigar, &ez->n_cigar, &ez->cigar);
}
kfree(km, mem2); kfree(km, off);
}
}
#endif

4
ksw2_extd2_avx.h
View File

@@ -35,6 +35,8 @@ Modified Copyright (C) 2021 Intel Corporation
#include <x86intrin.h>
#include <smmintrin.h>
#include <emmintrin.h>
void ksw_extd2_avx512(void *km, int qlen, const uint8_t *query, int tlen, const uint8_t *target, int8_t m, const int8_t *mat,
int8_t q, int8_t e, int8_t q2, int8_t e2, int w, int zdrop, int end_bonus, int flag, ksw_extz_t *ez);
void ksw_extd2_avx2(void *km, int qlen, const uint8_t *query, int tlen, const uint8_t *target, int8_t m, const int8_t *mat,
int8_t q, int8_t e, int8_t q2, int8_t e2, int w, int zdrop, int end_bonus, int flag, ksw_extz_t *ez);